XAI-ecDNA: A Proof-of-Concept Framework for Explainable Multi-Modal Extrachromosomal DNA Detection Using Synthetic FISH and Genomics

Abstract

Extrachromosomal DNA (ecDNA) is a key contributor to aggressive tumor evolution and drug resistance across multiple cancer types. Yet detection methods suffer from limited performance and lack interpretability. Moreover, paired fluorescence in situ hybridization (FISH) and genomic datasets necessary for developing multi-modal artificial intelligence tools remain unavailable in public repositories. We developed an explainable multi-modal framework combining genomic copy number profiles from 500 samples (TCGA, CytoCellDB) with synthetic FISH images generated via StyleGAN2-ADA trained on 247 published images. The framework employs early fusion of ResNet-50 image features and multi-layer perceptron processed genomic data through gated attention. A Biological Rationale Score (BRS) provides SHAP-based explanations with stability quantification and uncertainty estimates. Expert cytogeneticists validated synthetic image biological plausibility. Twelve clinical experts evaluated predictions in a crossover study. On synthetic images, our framework achieved AUROC 0.87, significantly outperforming genomic-only (0.72), image-only (0.79), and late fusion (0.84) baselines (all p<0.001). BRS identified MYC amplification (76% of cases), nuclear morphology (68%), and spatial clustering (56%) as top predictive features, aligning with known biology. Explanation stability reached Kendalls {tau}=0.74. Expert diagnostic confidence improved by 0.9 points (Cohens d=1.26, p<0.001) with explanations. Model calibration was excellent (ECE=0.042). This proof-of-concept demonstrates multi-modal explainable ecDNA detection potential. However, all results are on synthetic images; validation with real clinical FISH data is essential before deployment. The methodology provides a transparent template for interpretable AI development in data-scarce medical domains.